As large models continue to grow in parameter count, competition in AI infrastructure is shifting from raw computing power at individual nodes to system-wide coordination and efficiency. The server-centric model of building computing power now faces new challenges: even as GPU counts rise, inter-chip data exchange, power delivery, cooling efficiency and operational complexity are increasingly limiting how much computing power systems can actually deliver.

Against this backdrop, supernodes have emerged as a new focus for the industry. Rather than simply packing in more GPUs, they integrate computing, networking, power, cooling and software to turn one or more racks into a tightly coordinated computing system. At the 2026 Open Data Center Conference and inaugural Computing Expo, or ODX, Ruijie Networks showcased its Ethernet-based full-rack supernode solution, with a physical demo highlighting its rack-scale architecture, high-speed interconnects, device coordination and capacity for large-scale delivery.

To discuss supernode trends, technical approaches and industry strategy, the reporter spoke with Liu Yang, a supernode solutions expert at Ruijie Networks. Liu said a supernode should not be viewed simply as a server or networking product. At its core, it is a cluster system built for AI computing. By connecting multiple compute and network nodes through high-speed links, a supernode can improve overall processing efficiency and provide higher-performance support for large-model training and inference.

A supernode is not about “stacking GPUs,” but about turning the rack into a computer. In the era of large models, demand for computing power is changing. Traditional data centers have focused primarily on improving individual server performance and adding nodes through Scale-out architectures. But for models with hundreds of billions or even trillions of parameters, simply deploying more servers can reduce communication efficiency and resource utilization while driving up energy consumption.

A supernode, or SuperPoD, is a supercomputer that connects multiple computing accelerators, or XPUs, through a Scale-up network. Its core design uses high-speed interconnect protocols to tightly couple multiple compute nodes, creating a system that is logically unified but physically distributed, with high bandwidth, low latency and unified memory addressing. Liu believes the end state for supernodes is not a larger GPU cluster, but a system in which a rack—or even an entire row of racks—operates increasingly like a single computer.

In such a system, compute trays and switch trays will form new functional zones, redrawing the traditional boundary between servers and networks. From an industry perspective, supernodes are pushing AI infrastructure away from assembling individual devices and toward delivering complete systems. Future competition will depend not only on the performance of a single GPU, but also on who can integrate computing, networking, software and engineering delivery into a complex computing system that runs reliably.

For Ruijie, this direction draws closely on its long-standing expertise. The company has consistently focused on cluster solutions for AI data centers and is transferring its experience in Scale-out networking to Scale-up architectures. This allows it to deliver systems with lower latency, higher bandwidth, greater reliability and better support for XPU interconnects. Switch nodes are becoming a “second computing unit,” as networks begin to take part in computation—one of the biggest changes introduced by supernodes.

In traditional data center architectures, networking equipment and servers are usually supplied by different vendors, use separate interface standards and evolve independently. That works because network traffic in conventional workloads is relatively loosely coupled, with the network primarily responsible for data transmission. In the supernode era, however, switching capacity, physical form factor, processing capabilities and power budgets are all determined by GPU computing power and communication patterns.

Liu said switch and compute nodes in a supernode must be designed as an integrated system because they are closely linked at the physical, semantic and engineering levels. At the physical level, demand for high-speed interconnects requires coordinated placement of switching and computing components. As Scale-up networks target terabit-per-second bandwidth for each accelerator, copper interconnects face clear distance constraints: their effective transmission range is usually only one to two meters, and every added signal-boosting component increases power consumption and latency.

Switch-chip placement, GPU tray positioning and high-speed link design therefore need to be considered together. At the semantic level, switch nodes are also taking on more computing tasks. In collective communication workloads such as AllReduce and AllGather, relying entirely on GPUs consumes substantial computing resources and HBM bandwidth. In-network computing can offload part of this work to switch chips, allowing some operations to be completed as data passes through the switching fabric and reducing the burden of moving data between GPUs.

This means switch chips must be able to understand collective communication semantics rather than remaining limited to general-purpose packet forwarding. Communication libraries such as NCCL, HCCL and RCCL must also work closely with switch-scheduling algorithms so that the network can participate more efficiently in executing computing tasks. Liu said the switch node effectively becomes an accelerator with modest computing power—the meaning of the “second computing unit”—and its value must be measured by the effective computing power it delivers.

At the engineering level, power, cooling and fault domains must be designed as an integrated system. Total power consumption for a supernode rack is moving toward the megawatt range. Under this trend, power management is no longer an issue for individual components. It must be planned at the system level, coordinating power allocation among GPU accelerators, switch chips, CPUs, host boards and other components.

In a liquid cooling system, for example, parameters including coolant inlet and outlet temperatures and flow rates must be optimized against the combined thermal load of compute and switch nodes. System power consumption must also be dynamically balanced between computing and interconnect capacity. Supernode fault management similarly requires a unified system-level approach, with the supernode treated as a single unit for scheduling, management and fault tolerance.

Switch-link redundancy, fault-isolation granularity and RAS alert mechanisms must all work in concert with the GPU-side RAS framework. Otherwise, when a switch link fails, its recovery process may not align with the compute nodes’ fault-handling logic, undermining system stability and the supernode’s ability to sustain effective computing power over time. As interconnects move from copper to optical technology, supernodes are beginning to expand beyond their physical limits alongside the continued growth of AI models.

As a single rack expands from dozens of GPUs to 100 or more, traditional copper interconnects increasingly face limitations in bandwidth, distance and power consumption, making optical interconnects an important area of industry development. Liu said one key supernode trend is the continued expansion of the Scale-up domain. As rack capacity grows, communication-intensive parallel workloads place greater demands on interconnect performance. Because copper is constrained by physical distance, technologies such as optical backplanes and active optical cables are emerging as potential solutions.

Ruijie is also gradually incorporating optical interconnect architectures into its supernode roadmap. Its Supernode 3.0 solution targets Scale-up deployments for models with extremely large parameter counts. Using an optical interconnect architecture together with all-optical supernodes and NPO optical switches, it is designed to overcome the bandwidth bottlenecks of conventional electrical interconnects while supporting horizontal expansion of the Scale-up domain and computing clusters with as many as 512 accelerators.

Optical interconnects change more than the connection method; they also affect the design of the entire system. As high-speed optical modules move inside the rack, server layouts, switch-node designs and cooling methods all need to be reworked. For converged optical-electrical environments, Ruijie has developed end-to-end simulation and optimization capabilities. By refining component parameters and link designs, it aims to improve signal quality and system stability in optical interconnect deployments.

At the same time, supernode development is shifting from a singular focus on the scale of computing power toward overall system efficiency. Liu said supernode design is moving from being “compute-centric” to “memory- or bandwidth-centric.” In inference workloads especially, performance may depend more on HBM bandwidth and KV cache capacity than on peak computing power alone.

Supernode performance metrics are also shifting from MFU to tokens per second per watt or tokens per second per gigabyte. This means the next phase of AI infrastructure competition will focus more heavily on total system efficiency. Reducing data-movement costs, improving communication efficiency and extracting more effective computing power from every watt will become key measures of supernode performance. As AI racks enter a high-power era, power delivery and cooling are becoming architectural variables.

The immediate challenge created by rising computing performance is how to provide a stable power supply while removing heat fast enough. For supernodes, power and cooling are no longer supporting systems; they are architectural variables that shape the system itself. Power architectures must evolve in parallel. Liu said the industry is gradually moving from today’s 54-volt busbars toward technologies such as 800VDC high-voltage direct-current power and vertical power backplanes.

Raising voltage and reducing current can lower transmission losses while simplifying power distribution inside the rack. Cooling technology must also evolve, and cold-plate liquid cooling could eventually give way to more aggressive approaches. Liu said Ruijie has not simply opted for an all-liquid-cooling strategy. Instead, it weighs chip cooling requirements against commercial deployment costs. High-power components—including GPUs, CPUs, switch chips and retimers—use cold-plate liquid cooling to handle intensive thermal loads.

Other components, including capacitors, resistors, network interface cards and SSDs, continue to use air cooling, balancing thermal requirements with overall cost. This hybrid approach reflects a broader shift in AI infrastructure from a race for maximum performance to competition over engineering efficiency. For large-scale AI data centers, performance is only the foundation; reliability, deployment cost and operational complexity are equally important to commercial value.

Ruijie designed its supernode solution around coordinated power and cooling from the outset. In the RG-ETH 128 supernode solution, an integrated liquid cooling system supports a range of complex data center environments, provides ample power redundancy and can reduce overall facility energy consumption. Its modular 45OU rack supports flexible expansion of computing power, with peak system power delivery of up to 120 kW. As hardware-software integration becomes critical, Ruijie is building a complete AI computing ecosystem. Supernodes require more than hardware integration; the greater challenge is managing and using vast computing resources efficiently.

As GPU counts rise, fault management, resource scheduling, communication optimization and model compatibility all become important factors in determining how much computing power can actually be delivered. Ruijie is building a complete system spanning unified hardware management, a network operating system, an AI software stack and an operations platform. Around its supernode hardware architecture, the company has developed an in-house four-layer AI computing software stack. Through close hardware-software coordination, it aims to unlock the full performance of high-speed electrical and optical interconnects, liquid cooling and power delivery, creating a complete AI computing ecosystem from low-level hardware management to large-model applications.

Liu said the true technical barrier in supernodes is not building the equipment, but pushing system performance to its limits. The next phase will require advances in larger-scale Scale-up capabilities, coordination between computing and networking, and system-level energy optimization. The role of networking vendors is also changing. In the past, network equipment primarily provided connectivity. In the supernode era, networks must understand computing tasks, monitor model execution and participate in resource scheduling.

Ruijie aims to draw on its networking expertise to integrate computing, networking and management into a system-level AI infrastructure solution. The AI industry is entering a phase of deep infrastructure restructuring. The growth of model scale is driving more than demand for additional GPUs; it is forcing a redesign of the entire data center, from servers, networking, power and cooling to the software stack.

Ruijie’s supernode strategy is not simply an attempt to enter the server market. Instead, the company is using its networking expertise to extend high-speed interconnect capabilities deeper into AI computing. Through coordination between compute and switch nodes, hardware-software integration and unified electrical and mechanical design, it is exploring a new form of AI data center infrastructure for the large-model era. The ultimate goal is to combine computing, communication, power and cooling into one efficient system, converting every watt into more effective computing power.